TL;DR
- Since 2 August 2026, Article 50 of the EU AI Act requires a visible label on deepfakes: realistic AI images, video or audio that resemble real people, places or events.
- Improving a real photo is fine. Adding things that don’t exist is where it becomes misleading, and where a label is needed.
- AI-assisted editing, scripts, captions and ad copy don’t need a label. Fully automated publishing with no human review does.
- Personal profiles are not an exception, and YouTube, Meta and TikTok apply their own labels on top of the law.
- Fines for breaching the transparency rules go up to €15 million or 3% of global turnover.
“I use ChatGPT to make my product photos look more professional. They’re my real products. Do I have to say in the ad that the image is AI?”
I get this question in different forms almost every week, from clients, students and business owners. As a marketing strategist and GEO consultant who has worked with more than 300 brands and runs AI-assisted campaigns daily, I went through the AI Act, the Commission’s guidelines, the EU labeling icons and the platform policies, and turned them into one practical guide.
The short answer
Find your case
Open the group that matches what you are publishing.
Product and property photos
AI corrected light, color or sharpness on a real photoCosmetic edit, nothing changes about what is shown
No label
AI added features, objects or a setting that don't existThe image can mislead the buyer about reality
Label
An empty apartment "furnished" or extended with AIThe EU lists this as partially AI-modified content
Label
Background blur, upscaling, noise removalTechnical quality fixes
No label
People, faces and voices
A photorealistic AI person speaks or recommends in an adCan be mistaken for a real person
Label
Cloned voice or changed speech of a real personA classic deepfake
Label
Face swap or a heavily changed appearanceSubstantial change to a real person
Label
Beauty filter on your own videoMinor aesthetic edit
No label
Text, ads and books
Ad copy written with AI, on real visualsTreated as production help
No label
AI used for ideas, scripts, titles or subtitlesProduction help
No label
Article or book drafted with AI, then edited and signed by youHuman review and editorial responsibility
No label
Articles generated and published automaticallyNo human review, no editorial responsibility
Label
Political or social-issue ad with AI-altered visualsPlatforms require explicit disclosure
Label
Creative and personal content
Clearly fantastical image, such as a person riding a unicornNot realistic, so not a deepfake
No label
Parody, satire or an artistic project with realistic AIDisclose without spoiling the work
Light label
Realistic deepfake on a personal, non-commercial profilePublishing it online exposes it to the public
Label
What the law actually says
Article 50 of the AI Act splits responsibility between two roles:
| Role | Who that is | Obligation |
|---|---|---|
| Provider | Companies that build AI tools: OpenAI, Anthropic, Google, Midjourney | Mark AI outputs in a machine-readable way, such as watermarks or metadata |
| Deployer | Businesses, agencies, freelancers and creators who publish AI content | Visibly disclose deepfakes and certain AI-generated text |
The definition of a deepfake, from Article 3(60) of the AI Act: AI-generated or manipulated image, audio or video content that resembles existing people, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
Two conditions, both needed: resemblance to something real and an impression that it is authentic. This is why “it’s obvious to me that it’s AI” is not a defense. The test is whether it could mislead someone, not whether it misleads you.
For text, the rule covers AI-generated publications that inform the public on matters of public interest. It does not apply when a person has reviewed the text and holds editorial responsibility for it. The provider side of the same rules is why AI companies now add watermarks to AI text.
Key dates
- 10 June 2026Final Code of Practice on AI-generated content published
- 20 July 2026Commission guidelines on Article 50 published
- 2 August 2026Article 50 transparency obligations apply across the EU
- 24 September 2026Bulgarian AI bill submitted to Parliament, now in committees
- 2 December 2026Deadline for AI tools already on the market to add machine-readable marking
Where I draw the line, from practice
The rules say what. Here is how it looks in real work.
If AI only makes a real photo look studio-shot, no label. If it adds features the product doesn't have, label it, or better, don't do it.
I go to viewings where the online photos show furniture and views that aren't there. The EU uses exactly this case as an example of AI-modified content.
Late-night teleshopping with a "doctor" who isn't a real person recommending a product. This is what the rules were written for.
Fully automated programmatic SEO, generated and published with nobody reading it. That is where the text rule bites.
Use AI to help, then edit, check the facts and sign it. I tell my students the same: keep human involvement high. It protects you and makes the work better.
If you'd feel uncomfortable when a customer found out how an image was made, label it.
Personal profiles and platform rules
The label is required even on a personal profile with no commercial purpose. Publishing a realistic deepfake online exposes it to the public, which is no longer private use. And the platforms don’t distinguish between personal and business accounts:
| Platform | What it requires | If you don’t |
|---|---|---|
| YouTube | Disclose realistic altered or synthetic content when uploading. AI scripts, thumbnails and captions are exempt | Removal or Partner Program suspension for repeated failures. The label itself doesn’t limit reach or monetization |
| Meta | Ads made with its AI tools get an AI info label, next to “Sponsored” when there’s a photorealistic AI person | Meta can apply the label itself, including for third-party AI tools. Political ads need explicit disclosure |
| TikTok | Label realistic AI content. TikTok also labels AI content automatically when it carries C2PA metadata | TikTok can label it itself, and breaking its AI rules can lead to removal |
How to label correctly
The EU has published official icons: a basic AI icon, one for fully AI-generated content and one for partially AI-modified content. Using them is optional. Labeling is not.
- Visible from the start: the label must be clear when someone first sees the content.
- Not covered: no overlapping elements on top of it.
- Persistent: it should stay on when the content is reshared or downloaded.
- Platform settings: tick the AI disclosure on upload. Menus change often, so check the current version.
Detailed placement rules are in the Code of Practice, signed by about 190 companies and organizations by the end of July 2026.
Who is responsible and the fines
| Party | Responsibility |
|---|---|
| The business being advertised | Main legal responsibility to disclose, and the target of regulators’ fines |
| Agency or freelancer | Follows the rules on the client’s behalf. The business can seek compensation under their contract, and platforms can block the agency’s ad account |
| Platform | Provides labeling tools, and can label, remove or block content and accounts |
| AI tool provider | Machine-readable marking only, not how an advertiser later uses the content |
The fines under Article 99: up to €15 million or 3% of global annual turnover for breaching the transparency rules, with lower caps for small companies and start-ups. Banned manipulative practices go up to €35 million or 7%. The rules also apply to non-EU companies targeting people in the EU, and to EU businesses advertising abroad. In Bulgaria, a bill on the use and development of AI is still in committees, and the EU rules apply either way.
Your pre-publish checklist
Six questions, one minute
Why this matters beyond compliance
People notice. In research released by The Harris Poll at Cannes Lions 2026, 78% of consumers said AI makes ads feel less authentic, and 63% said they’d be less likely to buy from brands that lean on AI. In December 2025, McDonald’s Netherlands pulled its AI Christmas ad after the backlash.
So my rule is simple: use AI to make real things look better, not to show things that don’t exist.
Rules and platform settings change, so check the current sources for your specific case, and talk to a lawyer when the stakes are high. If you read the rules differently or have a case I haven’t covered, I’d like to hear it.
For a neutral overview with all the rules and sources, read EU AI Act Article 50, Explained on Egorithms.
About The Author
